def test_all_tasks_must_be_single_agent_or_multiagent(RPSTask, pendulum_task,
ppo_config_dict):
agent = build_PPO_Agent(RPSTask, ppo_config_dict, 'Test-PPO')
with pytest.raises(ValueError) as _:
_ = benchmark_agents_on_tasks(tasks=[RPSTask, pendulum_task],
agents=[agent],
num_episodes=1)
def test_if_populate_all_agents_is_set_a_single_agent_must_be_provided(
RPSTask, ppo_config_dict):
agent = build_PPO_Agent(RPSTask, ppo_config_dict, 'Test-PPO')
with pytest.raises(ValueError) as _:
_ = benchmark_agents_on_tasks(tasks=[RPSTask],
agents=[agent, agent],
num_episodes=1,
populate_all_agents=True)
def test_single_agent_can_populate_all_agents(RPSTask):
expected_winrates = [0.5]
actual_winrates = benchmark_agents_on_tasks(tasks=[RPSTask],
agents=[rockAgent],
num_episodes=200,
populate_all_agents=True)
np.testing.assert_allclose(expected_winrates, actual_winrates, atol=0.1)
def generate_evaluation_matrix(cool_game_params, logger):
# 0: SawBot 1: TorchBot 2: NailBot
benchmarking_episodes = 1
mcts_budget = 1
saw_vs_torch_task = generate_task('CoolGame-v0',
EnvType.MULTIAGENT_SIMULTANEOUS_ACTION,
botA_type=0,
botB_type=1,
**cool_game_params)
saw_vs_nail_task = generate_task('CoolGame-v0',
EnvType.MULTIAGENT_SIMULTANEOUS_ACTION,
botA_type=0,
botB_type=2,
**cool_game_params)
torch_vs_nail_task = generate_task('CoolGame-v0',
EnvType.MULTIAGENT_SIMULTANEOUS_ACTION,
botA_type=1,
botB_type=2,
**cool_game_params)
mcts_config = {'budget': mcts_budget}
mcts_agent = build_MCTS_Agent(saw_vs_torch_task,
mcts_config,
agent_name='MCTS agent')
saw_winrates = benchmark_agents_on_tasks(
tasks=[saw_vs_torch_task, saw_vs_nail_task],
agents=[mcts_agent],
populate_all_agents=True,
num_episodes=benchmarking_episodes)
nail_winrate = benchmark_agents_on_tasks(
tasks=[torch_vs_nail_task],
agents=[mcts_agent],
populate_all_agents=True,
num_episodes=benchmarking_episodes)
bench_msg = f'episodes={benchmarking_episodes} MCTS_budget={mcts_budget}'
winrates_msg = f'winrates=saw:{saw_winrates} nail:{nail_winrate}'
logger.info(bench_msg)
logger.info(winrates_msg)
logger.info(f'params={cool_game_params}')
return np.array([[0., saw_winrates[0], saw_winrates[1]],
[-saw_winrates[0], 0., nail_winrate[0]],
[-saw_winrates[0], -nail_winrate[0], 0.]])
def test_can_compute_winrate_for_player1_multiagent_task(RPSTask):
expected_winrates = [0, 1]
vs_paper = deepcopy(RPSTask)
vs_scissors = deepcopy(RPSTask)
# Ugly, would be awesome to have it in a one line
vs_paper.extend_task(agents={1: paperAgent})
vs_scissors.extend_task(agents={1: scissorsAgent})
actual_winrates = benchmark_agents_on_tasks(tasks=[vs_paper, vs_scissors],
agents=[rockAgent],
num_episodes=10)
np.testing.assert_array_equal(expected_winrates, actual_winrates)
def test_can_compute_cumulative_reward_for_agent_single_agent_task(
RPSTask2Repetitions):
vs_paper = deepcopy(RPSTask2Repetitions)
vs_scissors = deepcopy(RPSTask2Repetitions)
vs_paper.extend_task(agents={1: paperAgent})
vs_scissors.extend_task(agents={1: scissorsAgent})
expected_cumulative_rewards = [-2., 2]
actual_winrates, actual_cumulative_rewards = benchmark_agents_on_tasks(
tasks=[vs_paper, vs_scissors],
agents=[rockAgent],
keep_cumulative_rewards=True,
num_episodes=10)
np.testing.assert_array_equal(expected_cumulative_rewards,
actual_cumulative_rewards)